Untitled document

ciency in mimicking human language skills,

resulting in signi

cant transformations across various

nancial domain [10]. Applying LLMs to

sentiment analysis represents an innovative shift, where

the traditional sentiment analysis challenges are reinter-

preted and addressed through more advanced computational

approaches [11]. Their effectiveness is particularly notable in

tasks that require deep contextual understanding and nuanced

language interpretation, such as predicting market trends,

analyzing investor sentiments, and interpreting

Despite the growing interest in using LLMs for sentiment

news [10], [12], [13]

analysis, especially in

nancial markets, there remains a

cant gap in understanding the extent and nature of their

impact on

nancial instruments, particularly cryptocurren-

cies like Bitcoin. Existing literature predominantly focuses

on the technical capabilities of LLMs without adequately

exploring their practical implications in

analysis. Our study seeks to bridge this gap by not only

nancial sentiment

categorizing various LLMs and their applications in

markets but also by empirically investigating the correlation

between news sentiment, as processed by these models, and

Bitcoin price movements. This approach aims to provide a

more nuanced understanding of the role of media sentiment in

cryptocurrency markets. To achieve this, our study will

answer the following research questions (RQ):

1) RQ1: How does the classi

cation, data collection, and

application of LLMs in sentiment analysis in

uence

their effectiveness in

2) RQ2: What is the correlation between news sentiment,

nancial markets?

asanalyzedbyLLMs,andthepriceofcryptocurrencies

like Bitcoin?

This paper makes a signi

nancial sentiment analysis by integrating the advanced

cant contribution to the

capabilities of LLMs with the dynamic realm of Bitcoin

and cryptocurrency markets. The study stands out for its

comprehensive examination of various LLMs, including

BERT, FinBERT, and ChatGPT, within the speci

nancial sentiment analyses [12], [14], [15]. This area is

particularly challenging due to the nuanced language and

investor sentiments intrinsic to market dynamics [16]. The

systematic categorization and analysis of these LLMs in

the paper illuminate their individual strengths and collective

potential in enhancing

nancial market analytics. The focus

on the unique features and applications of these LLMs in

the

nancial domain reveals new insights into their

transformative role in market trend prediction and investment

decision-making.

By identifying a modest but discernible correlation

between news sentiment and Bitcoin prices, the paper

contributes valuable empirical evidence to the understanding

of cryptocurrency market dynamics. This insight is crucial

for a range of stakeholders, including investors,

analysts, and policymakers, who navigate the complexities

of these emerging markets. Additionally, the discussion of

challenges and future directions for LLMs in sentiment

analysis highlights both the current capabilities and potential

growth areas for these models in

Fig. 1. illustrates the process of Bitcoin transactions within a

nancial applications.

peer-to-peer network, showcasing how transactions are

signed and added to the blockchain through mining, while

also highlighting various factors that can in

This study is structured as follows: Section II presents the

existing literature closely related to sentiment analysis in

nancial markets. Section III outlines our research

methodology. Sections IV to VI are collectively focused on

addressing RQ1. Section IV presents a detailed classi

of LLMs in sentiment analysis. Section V discusses the data

cation

collection method and categorization. Section VI explores

the applications of LLMs in sentiment analysis. Section VII

provides a case study aimed at answering our RQ2, offering

practical insights into applying these models in a real-world

scenario. Sections VIII and IX discuss the challenges that

should be overcome when employing LLMs to solve sen-

timent analysis tasks and highlight promising opportunities

and directions for future research. The conclusions of our

study are presented in section X. The overall organization of

the paper is presented in Fig. 2.

Among the various LLMs, BERT (Bidirectional Encoder

Representations from Transformers) [14] has set a new

precedent in natural language processing by understanding

the context of a word in a sentence more holistically. BERT’s

architecture has been utilized in the

FinBERT [12], a model speci

nancial sector to create

ne-tuned to grasp the

subtleties of

nancial jargon and sentiment. FinBERT [12]

excels in interpreting complex

nancial reports, earnings

calls,andmarketanalysis,providingmoreaccuratesentiment

predictions than general-purpose models [17]. Additionally,

Ploutos [18], another

nancial LLM, demonstrates superior

performance in predicting stock movements. This model

uniquely integrates textual and numerical data using a

mixture of experts architecture, enhancing its ability to

deliver precise explanations for its predictions. A further

groundbreaking LLM is ChatGPT [15], which has been

instrumental in enhancing interactive

ChatGPT’s ability to engage in human-like conversations

nancial analysis.

and provide detailed, contextually relevant responses has

been utilized in customer service automation,

advisory, and real-time market analysis [19]. This model’s

sophisticated understanding of queries and ability to generate

coherent and context-aware responses make it an invaluable

tool in the dynamic world of

Few recent studies focus on various applications and

nance.

advancements of LLMs in

nancial sentiment analysis. In

a recent study, Sharma et al. [20] explored the use of

generative models like ChatGPT for sentiment analysis.

These models enhance sentiment analysis by augmenting

datasets with synthetic labeled data and simulating human

sentiment expression, particularly for tasks like sarcasm

detection.Keychallengesincludemaintainingthequalityand

consistency of generated data and addressing inherent biases.

By overcoming these issues, the potential for sentiment anal-

ysis in real-world applications can be signi

The architectures and applications of large language models,

cantly enhanced.

including their use in sentiment analysis presented by Raiaan et

al. [21]. They categorize different LLMs, such as GPT-3, and

explore their applications in various domains, including

nance. The paper also addresses the challenges and open

issues in deploying LLMs for sentiment analysis, such as

data scarcity and model interpretability. The increasing role

of generative AI models, such as GPT-3, in business and

nance is discussed in [22]. The work highlights the

potential of these models to generate realistic

perform sentiment analysis, and support decision-making

nancial data,

processes. The paper also explores the ethical implications

andregulatorychallengesassociatedwithusinggenerativeAI

Dong et al. [23] investigate the application of LLMs for

extracting relevant information from

The authors employ GPT-3 to analyze annual reports,

nancial documents.

earnings call transcripts, and other

nancial texts to identify

key sentiment indicators and predict stock price movements.

The study shows that LLMs can effectively process and

interpret large volumes of text data, providing valuable

insights for investors and analysts. Farimani et al. [24]

investigate the ef

ciency and accuracy of using LLMs like

GPT-3 for sentiment analysis in the

authors compare the performance of LLMs with traditional

nancial market. The

models,demonstratingsigni

cantimprovementsincapturing

nuanced sentiments and predicting market trends based on

nancial news and social media data. Another early review

[25] emphasizes the potential of both BERT and GPT-

2inadvancing

nancialsentimentanalysisthroughimproved

feature mapping techniques, leveraging their respective

strengths in understanding context and generating relevant

text.

While previous studies have signi

cial sentiment analysis using models like FinBERT and

integrated approaches combining sentiment indices with pre-

dictive models, our approach introduces a novel perspective

by leveraging more speci

c aspects like classi

collection and application with a case study. A central

cation, data

element of this research is the empirical investigation into the

correlation between news sentiment, as analyzed by LLMs,

and Bitcoin price movements. This case study is particularly

relevant given the growing in

uence of cryptocurrencies

like Bitcoin in global

nancial markets. Bitcoin serves as a

benchmark for the digital currency landscape, characterized

by its volatility, decentralized nature, and sensitivity to public

sentiment and news [26]. The study addresses the pressing

need to understand Bitcoin’s market behavior due to its

escalating impact on retail and institutional investors and its

potential in reshaping

nancial technology and monetary

This literature review adheres to the methodology proposed

by Kitchenham et al. [28], [29]. Following the guidelines

provided by Kitchenham et al. [28], our methods included

two main steps: planning and conducting the review.

Established academic databases were utilized to gather the

relevant literature, including Web of Science, IEEE Xplore,

Springer, arXiv, and UoA(University of Auckland) Library.

The following sections describe the methodology used to

source and evaluate the chosen literature. Speci

presents the structure of the literature review.

cally, Fig. 3

Our manual search encompasses four critical databases

knownfortheircomprehensivecollectionofscienti

The methodology involved a multi-step process, beginning

cpapers.

with creating a keyword dictionary instrumental in the initial

search across these databases. Our search string should

combine two sets of keywords: one related to sentiment

analysis and the other to LLMs. If the paper contains both

types of keywords, it is more likely that it is the paper we

need. The complete set of search keywords is as follows

:

detection, Opinion mining, Emotional analytics, Affec-

tive computing, Polarity classification, Subjectivity

analysis, Sentiment scoring, Mood analysis, Opinion

polarity, Sentiment quantification, Emotion recogni-

tion, Tone analysis, Sentiment lexicons, Sentiment

metrics, Textual affect detection, Semantic orien-

tation, Sentiment strength, Sentiment benchmarks,

Sentimental analysis tools, Review analysis, Con-

sumersentiment,Investorsentiment,Marketsentiment,

Brand sentiment, Social sentiment, Sentiment cor-

relation, Aspect-based sentiment analysis, Sentiment

summarization,Sentimentalclassification,Sentimental

interpretation.

Large Language Model, Pre-trained, PLM, Pre-

training,NLP,NaturalLanguageProcessing,DL,Deep

Learning, ML, Machine Learning, ChatGPT, Neural

Network, Transfer Learning, Sequence Model, T5,

GPT, Codex, BERT, Transformer, Attention Model, AI,

Artificial Intelligence.

We included keywords like Machine Learning and Deep

Learning, alongside other terms not directly related to Large

Language Models (LLMs), in our search criteria. This

broader approach aims to ensure we don not overlook any

relevant research, thereby expanding our search scope during

automated searches.

Upon retrieving these papers, the next step involved

a detailed examination of the titles and abstracts, which

allowed to determine the relevance of each paper to

the research objectives based on inclusion and exclu-

sion criteria. We designed these criteria following several

several state-of-the-art papers [30], [31], as shown in

Table 1, so that the selected documents can directly address

our topic. We drop duplicated studies across multiple

databases to re

ne our dataset, streamlining our literature

Following the curation of unique studies, we proceeded to

a more in-depth review, scanning the full text of each

selected paper. A thorough quality assessment can help

mitigate biases that may arise from low-quality studies

and guide readers on where to approach conclusions with

caution [32]. We developed a set of ten Quality Assessment

Criteria (QAC), detailed in Table 2. These criteria evaluate

the papers’ relevance, clarity, validity, and importance.

The

nal stage in our search process was to conduct a

quality assessment of these primary studies, evaluating

them against prede

ned criteria to ensure that only the

most rigorous and relevant research was included in our

analysis.

The systematic literature review on LLMs for sentiment

analysis acknowledges the risk of missing key studies due to

potential gaps in keyword summarization. To mitigate this,

a dual approach combining manual review and auto-mated

searches was utilized, with keywords derived from

authoritative sources and forward and backward snowballing

techniques employed to ensure thoroughness. Additionally, to

counter study selection bias, de

exclusion criteria were established, and a QAC framework

ned inclusion and

was implemented, with ambiguous cases receiving manual

scrutiny. This blend of strategies aimed to balance ef

with meticulousness, reducing biases and enhancing the

ciency

This section explores the classi

contextofsentimentanalysis,emphasizinghowtheirsizeand

cation of LLMs in the

architecture impact their effectiveness. We categorize LLMs

based on their structural design, distinguishing between

encoder-only, encoder-decoder, and decoder-only models,

each with distinct capabilities in processing natural language.

For an in-depth exploration of the relevant literature, we have

included a comprehensive summary in Table 4.

Pre-trained language models (PLMs) have proven highly

effective in various natural language processing (NLP) tasks,

as evidenced in several studies [33], [34]. Researchers have

noted that increasing the size of these models signi

boosts their capabilities, particularly when the size of

cantly

parametersexceedsacertainpoint[35],[36].Thedesignation

‘‘Large Language Model’’ (LLM) is used to differentiate

language models based on their size, primarily referring to

PLMs with a larger scale [37]. However, it is important to

mention that there is no widely agreed-upon standard in the

literature for the minimum parameter size for an LLM, as

its ef

ciency is linked to the dataset’s size and the total

computing power used. In our study, we follow the

classi

cationandtaxonomyofLLMsintroducedbyPan et al.

[38], dividing mainstream LLMs into three categories based

on their architecture: encoder-only, encoder-decoder, and

decoder-only. This classi

cation and the corresponding

models are depicted in Fig. 3.

Encoder-only LLMs are a speci

framework that employs solely the encoder part of the model.

c type of neural network

The primary role of the encoder is to process and transform

theinputtextintoahiddenrepresentation.Thisrepresentation

is critical in understanding the connections among words

and the general context of the sentence. Prominent examples

of encoder-only LLMs include BERT [14] and various

adaptations of it [12], [39], [40]. BERT, in particular, is built

on the encoder architecture of the Transformer [41]. Its

unique feature is the bidirectional attention mechanism,

which allows it to analyze the context to the left and

right of each word concurrently during its training phase. In

the

nancial domain, other prominent models like

FinBERT[12],CryptoBERT[42],andSBERT[26]havebeen

widely employed.

These models distinguish themselves from the original

BERT [14] by enhancing the architecture to include novel

pre-training tasks or adjusting to different data modalities,

thereby improving their effectiveness for

BERT [14] that is pre-trained explicitly on

ne-tuned to perform sentiment analysis within the

nancial domain, achieving an accuracy of 0.86 and an F1-

Score of 0.84. Similarly, CryptoBERT [42], which is also

grounded in the BERT [14] model, undergoes

ne-tuning on

a cryptocurrency-speci

c corpus, yielding heightened

accuracy in the sentiment classi

cation of texts related to

cryptocurrencies. It achieved accuracy scores of 55.60 and an

F1-Score of 55.79 among

ve models for the StockTwits

1

data,

which contains 1.875 million posts. These models have

demonstrated their pro

ciency in various applications, such

as predicting market movements, analyzing investor

sentiments, and automating

nancial report summaries,

showcasing their transformative impact on the

Encoder-decoder LLMs integrate both the encoder and

decoder components [41]. The encoder component converts

the input text into a hidden representation, adeptly grasping

the fundamental structure and meaning. This con

representation is a transitional language, facilitating the

dential

connection between input and output formats. On the other

hand, the decoder leverages this hidden representation to

produce the desired output text, transforming the abstract

representationintospeci

c,contextuallyappropriatephrases.

Within this context, the memory module of models like

FINMEM [43] stands out. It mirrors human cognitive

processes, providing clear interpretability and

real-time adjustments. This feature enhances the model’s

exibility for

utility in

nancial trading by allowing it to hold on to

essential information for extended periods, which is crucial

for complex decision-making. FINMEM outperformed in

trading

ve different stocks, achieving the highest Sharpe

ratio of 2.6789 and lowest max drawdown of 10.7996%.

Another example is TradingGPT [44], an innovative LLM

multi-agent framework endowed with layered memories. The

ability of TradingGPT to navigate through

https://stocktwits.com/

its application in trading exempli

LLMscanbepotenttoolsinenhancingtradingstrategies[44].

es how encoder-decoder

Decoder-only LLMs exclusively use the decoder module to

produce the intended output text. They follow a unique

training approach focusing on sequential prediction [45].

Contrary to the encoder-decoder framework, where the

encoder handles the input text, the decoder-only structure

starts from a base state and sequentially predicts tokens,

thereby progressively constructing the output text. This

method heavily depends on the model’s pro

grasping and predicting language structure, syntax, and

ciency in

context. Key examples of this architecture include the GPT

series models such as GPT-1, GPT-2, GPT-3, GPT-4, and

[45],[46],[47],[48].The

GPT series has shown promising performance in

sentimentanalysis,notonlyforTwitternewsbutalsointerms

of accuracy, recall, and F1-score across different forex pair

news [49], [50]. These models demonstrate their capability to

excel in

nancial contexts, highlighting their potential for

improving sentiment analysis and market prediction tasks.

These models can execute downstream tasks with min-

imal input, often requiring just a handful of examples or

straightforward instructions. This attribute eliminates the

need for additional prediction heads or extensive

processes, rendering them particularly valuable in sentiment

ne-tuning

analysis research. For instance, recent developments in the

industry have witnessed Google unveiling Bard. At the

same time, Meta has introduced its models, LLaMA [51]

and LLaMA2 [52], alongside Microsoft’s foray with Bing

Chat.

3

One application of LLaMA in the realm of

sentiment analysis is demonstrated by FinMA, a version of

ne-tuned for this task, which recorded

the highest F1-score of 0.87 on the FiQA dataset [53].

Furthermore, LLaMA2 has proven effective, reaching an

accuracyof84.03%throughsupervisedlearningandaligning

nancial texts [54]. These developments highlight the capa-

bilities of LLaMA models in sentiment analysis, particularly

their pro

ciency in the precise interpretation and assessment

This section examines the methodologies employed in

collecting and utilizing datasets for sentiment analysis in

LLMs. This section underscores the pivotal role of data in

training LLMs, emphasizing the need for diversity and

comprehensiveness in dataset collection to enhance model

performance in varied contexts [55]. We explore the

systematic process of dataset categorization, preprocessing,

and formatting, which is essential for aligning data with the

model’s training objectives and processing needs.

2

https://chat.openai.com/

3

https://www.microsoft.com/en-us/edge/features/bing-chat

Data is a vital and essential component in training Large

Language Models (LLMs), signi

generalization capabilities, ef

ciency, and overall perfor-

mance [55]. An ample amount of high-quality and varied

dataenablesmodelstothoroughlylearnfeaturesandpatterns,

ne-tunetheirparameters,andmaintaindependabilityduring

validation and testing.

Our initial focus is on examining the methodologies for

dataset acquisition. Through this analysis of data

collection techniques, we have categorized the sources of

data into four groups: open-source datasets, datasets that

are actively collected, datasets that are speci

constructed, and datasets derived from industrial sources.

cally

Open-source datasets [56], [57] are publicly available data

compilations typically distributed via open-source platforms

or repositories. An example of this is the FiQA [56] dataset, a

substantial new dataset featuring Question-Answering pairs

focusedon

nancialreportscraftedbyexpertsin

credibility of these datasets is bolstered by their open-source

nance.The

status, enabling community-based updates and ensuring their

reliability for scholarly research.

The Financial PhraseBank,

rst introduced by Malo et al.

[58], consists of 4,845 English sentences randomly selected

from

nancial news articles in the LexisNexis database.

These sentences were annotated by 16 experts in

business who evaluated how the information could in

nance and

the stock prices of the companies discussed. Furthermore,

uence

the dataset includes information about the level of agreement

among the annotators regarding the sentiments expressed in

nancial is a specialized subset of the TRC24

4

collection from Reuters, which encompasses 1.8 million

news articles released between 2008 and 2010. This subset

speci

cally contains 46,143 documents, totaling nearly 29

million words and close to 400,000 sentences [12].

SemEval 2017 Task 5 focuses on

analysis (FSA) of news headlines and microblogs [59].

ne-grained sentiment

The training set for this task includes 1,142

headlines and 1,694 microblog posts, each annotated with

nancial news

target entities and their corresponding sentiment scores.

The test set comprises 491

nancial news headlines and

794 posts [11].

Collected datasets [26], [60] are compiled by researchers

from diverse sources, such as signi

blogs, and social media. Researchers often extract data from

cant websites, forums,

sources like Twitter and Reddit for datasets speci

tailored to their research inquiries.

cally

Constructed datasets [12] are researcher-generated

datasets derived from modifying or enhancing collected

datasets to align with speci

c research goals. Manual or

semi-automatic modi

cations can include creating domain-

speci

c tests, annotated datasets, or synthetic data.

4

https://trec.nist.gov/data/reuters/reuters.html

Industrial datasets [10] sourced from commercial or

industrial entities contain proprietary data and are essential

for research addressing real-world business contexts.

Acknowledging that certain studies utilize diverse datasets

encompassing various categories is essential. For instance,

Wuetal.[10]trainedBloombergGPTusingmultipledatasets,

e.g., complex table datasets and question-answering pairs.

The data types are crucial in determining the architecture

and choice of LLMs, as they directly affect the extraction of

implicit features and the decisions made by the model [67].

The selection of speci

uence the LLMs’ overall effectiveness and ability to

c data types can signi

generalize. In our research, we explore and categorize the

various types of

nancial datasets used in studies of LLMs

for sentiment analysis. By examining how data types relate to

model architectures and their performance, we aim to

highlight the importance of data types in the effectiveness of

LLMs for sentiment analysis.

We classi

ed the data types of all datasets into

categories: Twitter posts, Reddit posts, News articles, Annual

ve

reports, and Fi-QA. Table 3 describes the speci

included in the data types corresponding to the datasets we

c data

summarized from the 15 studies.

This section delves into LLMs’ diverse and transformative

applications in

nancial sentiment analysis. In recent years,

integrating advanced LLMs into the

cant evolution in how

trends, and investor sentiments are analyzed and interpreted.

nancial data, market

This section explores how LLMs predict market trends,

optimize trading strategies, and forecast stock prices.

This section explores the application of LLMs for predicting

cryptocurrency market trends, with a particular focus on inte-

grating sentiment analysis into these predictions. The poten-

tial of LLMs to distill sentiment from vast datasets offers a

novel dimension to the forecasting models, as evidenced by

several recent studies. Zou and Herremans [61] intro-duced

a pioneering multimodal model, PreBit, speci

cant Bitcoin price movements.

Bashchenko [26] provided insights that counter the notion of

Bitcoin’s value being purely speculative, demonstrating that

non-endogenous news carries fundamental information

affecting Bitcoin prices.

Raheman et al. [62] highlighted the practical advantages of

interpretable AI and NLP methods over non-explainable

alternatives, suggesting that transparency in AI could lead to

more valuable applications in the

Lessmann [68] demonstrated the advantages of re

nancial sector. Ider and

FinBERT with weakly labeled data, illustrating how even

ning

imprecisely labeled datasets can signi

based feature prediction and forecasting accuracy for cryp-

cantly improve text-

tocurrency returns. Their study utilized a dataset comprising

433 test samples, with a noteworthy agreement rate of 92.6%

among all 16 expert labels. This approach facilitated the

development of predictive models for Bitcoin and Ethereum

that substantially outperformed baseline models, achieving

gains of 0.572 and 0.501, respectively.This evidence under-

scores the ef

cacy of leveraging weak labels in enhancing

the performance of

nancial prediction models, particularly in

the volatile domain of cryptocurrency markets.

Ortu et al. [63] investigated cryptocurrency price pre-

diction by analyzing social sentiment data from GitHub

and Reddit, employing a pre-trained BERT-based model to

synthesize emotional and sentiment indicators from social

media commentary into hourly and daily series datasets.

Their

ndings indicated that incorporating these social

sentiment metrics markedly enhances the predictive accuracy

for the daily pricing of Bitcoin and Ethereum. The research

highlights a signi

cant inverse relationship between negative

sentiment and price volatility within the Bitcoin market, sug-

gesting that users might interpret volatility as a speculative

opportunity. In contrast, the Ethereum market sentiment is

predominantlyin

uencedbyemotionalarousal,whichshows

a substantial positive correlation with negative sentiment,

indicating that community reactions are more emotionally

driven rather than directly related to price movements.

Building on these

ndings, Nguyen et al. [27] explored the

distinctive impact of ChatGPT-based sentiment indicators on

Bitcoinreturns,revealingitsadeptnessatsentimentdetection.

The study’s results prompt further investigation into how

Generative AI might enhance

nancial data analysis and

social media sentiment interpretation, potentially unlocking

more sophisticated market insights. This research opens up

new pathways for sentiment analysis in

leveraging AI technologies.

nancial markets,

Developing a robust trading strategy is crucial in the volatile

realm of

nancial markets, where integrating sentiment

analysisandLLMscanprovideacompetitiveedge.Kim et al.

[60] leveraged an LLM adapted to the crypto domain to

parse crypto news sentiments called CBITS. Their

research demonstrates that trading strategies augmented

with sentiment scores signi

cantly outperform conventional

models, underscoring the ef

cacy of sentiment-based trading

approaches. Backtesting various Bitcoin trading strategies,

their study reveals that models employing TabNet combined

with RoBERTa, speci

cally the TabNet RoBERTa top 10,

yield the highest pro

t, recording an impressive gain of

304.65%. In contrast, other models assessed during the same

test period generated negative returns.

Yu et al. [43] introduced FINMEM, an innovative LLM-

based framework crafted for

This framework is structured around three central modules:

nancial decision-making.

Pro

ling, which tailors the agent to speci

Memory, which processes

nancial information in a layered

mannerakintohumancognitivestructures,facilitatingdeeper

assimilation of

nancial data; and Decision-making, which

translates the processed information into actionable invest-

ment strategies. The adaptability of FINMEM, particularly

its memory module, provides a level of interpretability that

mirrors human trading logic, coupled with the capability for

real-time adjustment to optimize trading decisions.

Li et al. [44] took the concept further by developing an

LLM multi-agent framework with layered memories called

TradingGPT. The LLMs at the heart of this framework act as

decision-making cores for trading agents, utilizing the

layered memory system to synthesize historical data and

current market conditions. This innovative approach enables

the agents to engage in strategic dialogues with peers, re

their investment choices, and uphold a diverse yet robust

ne

decision-making process informed by their unique trading

personas.

Curtó et al. [69] provided empirical evidence showcasing

the adaptability of LLM-informed strategies to the dynamic

bandit problem, a standard paradigm in trading strategy

formulation. Their experiments underscore the ability of

LLMs to navigate the complexities of the

yielding a strategy that competes favorably with traditional

nancial markets,

methods even in unpredictable scenarios.

Gupta [65] aimed to streamline the analysis of Annual

Reports across various

rms by harnessing the analytical

prowess of LLMs. A machine learning model was trained

using these insights as predictive features by distilling

insights from the LLMs into a quantitatively styled dataset

and supplementing it with historical stock prices. The

walk-forward testing indicated that such a model could sig-

ni

cantly outperform benchmarks like the S&P 500 returns,

underscoring the potential of GPT3.5 to revolutionize trading

strategies.Theresearchrevealedthatthemodel,whenusedto

select the top k stocks, consistently generated higher returns

than the S&P 500. Notably, the returns were inversely related to

the value of k, with lower k values correlating with higher

returns. This outcome indicates that the stocks predicted as

top performers by the GPT model indeed yielded better

Our analysis underscores the broad utility of LLMs in stock

price prediction through sentiment analysis, showcasing

their versatility across various

nancial applications. Araci

introduced FinBERT, a model tailored for the

sector, demonstrating superior capabilities in economic text

miningandsuggestingfurtherapplicationofFinBERTacross

different

nancial NLP tasks. FinBERT’s utility could be

cantly extended by integrating more extensive stock

market datasets, presenting opportunities for more intricate

market analysis and model re

Mishev et al. [11] provided evidence that contextual

nement.

embeddings substantially improve ef

analysis over traditional lexicons and static word encoders, a

ciency for sentiment

bene

t that holds even in the absence of large datasets. This

advancement points to the potential of LLMs to revolutionize

sentiment analysis with a more profound understanding of

contextual nuances in

Deng et al. [66] revealed that LLMs can achieve remark-

nancial texts.

able outcomes in market sentiment analysis. The study

showed that with minimal examples, it is possible to calibrate a

‘student’ model that matches or surpasses the performance of

more extensive, state-of-the-art models, optimizing both

effectiveness and computational ef

Fazlija and Harder [64] identi

ciency.

derived from news content play a critical role in predicting

ed that sentiment scores

the direction of stock prices. The correlation between

news sentiment and market performance underscores the

value of high-quality, content-based sentiment indicators in

This case study aims to explore the relationship between the

sentiment expressed in cryptocurrency news articles and the

price

uctuations of Bitcoin. Leveraging the power of

sentiment analysis through advanced language models, this

study seeks to provide a deeper understanding of how public

sentiment, as re

ected in media [16], can impact

markets, particularly the volatile cryptocurrency sector.

We collected comprehensive daily cryptocurrencies data

from the investing website, www.investing.com,

5

to inves-

tigate this relationship. The dataset spans two years, from

November 1, 2021, to November 1, 2023, and encompasses

various metrics, including price, closing price, highest and

lowest price of the day, opening price, and volume of

transactions. Consistent with methodologies employed in

targetvariable.ThisdecisionisbasedontheBashchenko[26]

price’s everyday use as a critical indicator of market

sentiment in

nancial research, providing a reliable measure

of the market’s end-of-day valuation for Bitcoin.

Given the current existence of approximately 1,000

cryptocurrency coins, some of which suffer from incomplete

information or delayed publication, our selection criteria

focused on coins with at least 1,000 recorded observations.

This threshold ensures the accumulation of suf

our analyses. Importantly, the use of transfer entropy as a

cient data for

methodological approach in our study offers the advantage of

not necessitating a balanced dataset, thus allowing for a

broader inclusion of data points. Our dataset represents over

80% of the cryptocurrency market’s total market

capitalization, ensuring a comprehensive analysis scope.

To gather cryptocurrency-related news data, we employed an

open-source Python library, scrape,

ciency in web scraping. This tool was instrumental in

compiling a substantial dataset of tweets about various

cryptocurrencies. To ensure a targeted and relevant data

collection, we used a set of carefully selected search

keywords for each cryptocurrency. For instance, in the case of

Bitcoin, the search parameters included a combination of its

name and symbol, such as ‘BTC OR BTC OR BITCOIN OR

Bitcoin’. Aligning with the timeframe of our price data, the

collection period for the news data was also set from

November 1, 2021, to November 1, 2023.

Given the sheer volume of cryptocurrency-related news

and the constraints of our computational resources, it was

necessary to limit the quantity of news collected daily for

each cryptocurrency. Without such a limit, the sentiment

analysis process would have been impractically prolonged,

5

https://www.investing.com/crypto/bitcoin/btc-usd

6

https://github.com/JustAnotherArchivist/snscrape

potentially taking months. To achieve this, we implemented a

timed request mechanism, where news was requested every

20 seconds using a single computer. This approach was

crucial to avoid triggering anti-crawler mechanisms on the

websites we scraped, while also ensuring a consistent data

collection rate. By limiting requests in this manner, we

capped the collection at a maximum of 5000 news articles per

day for each cryptocurrency, totaling 18506 articles for the

entire study period. For each piece of news, we

meticulously recorded several key attributes: the date and time

of the post (DateTime), the headline, the main text of

thenewsarticle,theauthor’sinformation,theURL,andafew

other relevant features.

FinBERT, introduced by Araci in 2019 [12], stands as the

c BERT model, pretrained on

the expansive TRC2-

nancial corpus. This corpus, a spe-

cialized subset of Reuters’ TRC2, comprises approximately

1.8 million news articles published between 2008 and 2010.

Given the scope of this paper, a detailed exploration of the

BERT architecture is beyond our purview, but readers are

encouraged to consult Araci’s original work for a

comprehensive understanding.

The FinBERT model underwent further

the Financial Phrase Bank, a resource developed by Malo et

ne-tuning using

al. in 2013 [71], speci

cally for sentiment classi

nancial domain. FinBERT’s performance in

nancial sentiment analysis tasks showed a notable 15%

improvement over generic BERT models [12]. This

enhancement in accuracy and the successful application of

FinBERT in studies parallel to ours, such as those by Zou

and Herremans [61] and Farimani et al. [72], underscored its

suitability for our research objectives.

For our study, we opted to employ the FinBERT model in

its pre-

ne-tuned state, initially con

gured by Araci [12]. This

decision was driven by the nature of our data set, which

primarily consists of unlabeled news articles. Further

tuning of FinBERT on other labeled datasets was deemed

ne-

unnecessary, considering it has already been optimized for

sentiment classi

cation using the Financial Phrase Bank. By

applying this ready-to-use,

nely tuned FinBERT model to

our news data, we aimed to leverage its advanced

capabilities for accurate sentiment analysis in the

sector without additional training. FinBERT outputs

sentiment scores for each news article on a scale from 1 to

10, where 10 indicates a high con

thenewspositivelyimpactingBitcoinprices.Finally,thedata

dence level in

are stored in a database (Fig. 4).

Engle’sbivariateDynamicConditionalCorrelation-Extended

Generalized Autoregressive Conditional Heteroskedastic-

ity (DCC-GARCH) technique, introduced in 2002, is a

cornerstone model for analyzing the co-movement of

cial returns. This technique boasts two signi

nan-

over other variants in the GARCH model family, such as the

cant advantages

Baba-Engle-Kraft-Kroner(BEKK)andConstantConditional

Correlation (CCC) models. Firstly, it exhibits a superior

capacity to capture time-varying conditional covariance. This

is achieved with less computational complexity than the

BEKK model, making it more ef

complex analyses. Secondly, unlike the CCC model, which

cient and accessible for

assumes constant correlations over time, the DCC model

allows for variation, adding

exibility and realism to the

analysis.

The bivariate DCC model’s simplicity is particularly

bene

cial in many return series contexts. A key strength is

its ability to directly account for heteroscedasticity by

calculating Dynamic Conditional Correlations (DCCs) from

standardized residuals. As Chiang et al. [73] noted, this

approach ensures that the DCCs are free from biases

associated with volatility clustering, addressing concerns

highlighted by Forbes and Rigobon [74]. Additionally, the

DCC model’s pro

ciency in generating accurate, time-

varyingestimatesofvolatilitiesandcorrelationsisinvaluable.

It capably re

ects the latest market news and responds to

regime shifts triggered by shocks and crises. This dynamic

analysis of correlations over time facilitates more informed

asset allocation and hedging decisions.

Implementing the DCC model involves a two-step process

to ascertain conditional correlations. A univariate GARCH

model is initially estimated for each return series, yielding

the conditional variance. Subsequently, dynamic conditional

correlations are derived from these standardized residuals.

Following the methodology described by Bauwens and

Laurent, the model is delineated like this

where the return vector R

and sector indices, r

S

and the selected alternative investments, r

theconditionalmeanprocess,andZ

independent identically distributed random variables vector.

The conditional covariance matrixt

conditional correlation matrix

denotetheunivariateGARCHvariances,

The (2

×

2) symmetric positive matrix Q

where C is the unconditional correlation matrix of standard-

ized innovations

η

t

, The added value of the positive scales

1. We obtain the DCCs by

Transfer entropy offers distinct advantages over tradi-

tional methods, enhancing its capability to evaluate infor-

mation

ows, as highlighted by Barnett et al. [75].

Unlike conventional econometric models that rely heavily on

domain-speci

c assumptions and constraints, transfer

entropy facilitates a non-parametric analysis of time-series

data, minimizing the need for extensive presumptions about

stochastic processes. Fundamentally, transfer entropy is

grounded in econophysics, focusing on quantifying the

directional information

ow of a variable over time, rooted

ininformationtheory.Thisconceptwasoriginallyintroduced

In this context, i denotes a discrete random variable

characterized by its probability distribution, p(i), re

the various outcomes it may manifest. H is identi

ecting

most effective function for facilitating this transformation,

ed as the

and H

I

is known as Shannon entropy. Shannon’s [76]

seminal work in 1948 established the groundwork for this

methodology, focusing on the uncertainty and dynamism in

a variable’s processes. Subsequently, Kullback and

Leibler [77], in 1951, expanded upon this by integrating an

additional element, referred to as process J. Notably, the

concept of Transfer entropy gains complexity with the

inclusion of more variables and values, indicating a broader

and more intricate understanding of entropy.

To elaborate further, the marginal probability distributions

p(i)

,

p(j) and the joint probability distribution p(i

,

j) are

expected to form a stationary time series. This implies that

) represents a sequence of values over

ned for process J in a comparable

manner. Kullback and Leibler [77], in their 1951 work,

introducedabroaderapplicationoftheMarkovprocesstothis

Transfer entropy revolves around the likelihood of one

variable obtaining information from its past and from another

variable (j

t

). This core idea behind ‘Transfer entropy’ is to

quantify the information exchange between two distinct,

random variables. Schreiber [78] elucidated this approach,

where I and J represent two separate processes. The formula

for transfer entropy from J to I is de

between the information absorbed by a future instance of

ned as the difference

from the past values of both I and J, and the

information absorbed by the same future instance solely from

the past values of I. In essence, transfer entropy seeks to

measure the net information

is used to assess the

information from J to I. Dimp

ow of

and Peter [79] introduced

novel methods, including the Markov block bootstrap and

the repeated bootstrap, to this

eld of study. They base their

investigation on the null hypothesis which posits the absence

of any information transfer.

Here, J and I represent two distinct processes, while q is a

positive weighting parameter q

>

0 applied to the individual

probability function p(

.

) for computations. Speci

element of the time series I, and j

n

denotes

then6thelementofthetimeseriesforthevariableJ.Itshould

escort distribution as de

purpose of introducing the Markov process into this analysis is

to estimate the likelihood of transitioning from one state to

anotherduringinformationtransfer,aswellastofacilitatethe

prediction of potential transition matrix scenarios. According

to Equation 10, and following the methodology suggested by

Bekiros et al. [80], setting l

denoising of the dataset and enables Transfer entropy to

detectasymmetricalinteractionsbetweenpairs(X andY)and

(Y and X), thus offering valuable insights into the dynamics of

information

ow between two time series. In essence,

transfer entropy relies on the logarithmic scale of the number

of possible outcomes, determined by a given probability

distribution, to analyze information

Table 5 presents the descriptive statistics for a case study

analyzing the volatility of Bitcoin prices and news sentiment

from November 1, 2021, to November 1, 2023. This

period witnessed signi

uctuations in Bitcoin prices, as

evidenced by a minimum price of $15,766 on November 21st,

2022, and a peak of $67,526 on November 8, 2021. Brown

[81] suggests that kurtosis values typically range from

10, and skewness values between

acceptable. In this context, the Bitcoin price exhibits a

skewness more signi

cant than one and a kurtosis exceeding

3, indicating a distribution with a higher peak and thicker

tails than a normal distribution, thereby implying a higher

likelihood of extreme values. Regarding news sentiment,

the skewness is 0.9311, denoting a moderate positive skew

with a longer right tail. The kurtosis of 4.6131, above 3,

categorizes the distribution as leptokurtic,

suggestingthatthenewssentimentscoreshaveasharperpeak

and heavier tails compared to a normal distribution.

Table 6 details the outcomes of unit root tests conducted on

the variables utilized in this study, explicitly presenting the

Augmented Dickey-Fuller (ADF) test results for each factor.

Despite the fact that stationarity is not a prerequisite for

utilizing the transfer entropy approach, which can handle

probability density functions from a single realization as

highlighted by Wollstadt et al. [82], we nevertheless

proceeded to perform a stationarity test.

For the Bitcoin price, a

rst-order difference was applied.

The ADF test result for Bitcoin price, with a statistic of

2.4729 and a higher p-value, indicates that the time series is

non-stationary. This means that the null hypothesis of a unit

root for Bitcoin price cannot be rejected at the 5%

cance level, necessitating further analysis due to its

non-stationary nature.

In contrast, the ADF test for News Sentiment yields a

statistic of

6.0595 with a p-value of 0.01. This p-value,

being below the commonly accepted signi

0.05, strongly refutes the null hypothesis of a unit root.

cance level of

Consequently, we can con

dently reject the null hypothesis,

af

rming that the news sentiment time series is stationary.

Other cryptocurrencies show mixed results: BNB has an

ADF test statistic of

2.9179, indicating non-stationarity as

the null hypothesis cannot be rejected. ETH has an ADF

test statistic of

2.4665, which is non-stationary. DOGE

showsastatisticof

3.7383,whichrejectsthenullhypothesis

at the 5% signi

cance level, indicating stationarity. TRON

has an ADF test statistic of

stationarity. XRP has an ADF test statistic of

2.7859, indicating non-

indicating non-stationarity. SOL and ADA show statistics of

2.8493, also

2.8546 and

4.3383 respectively, indicating stationarity..

These results suggest that while News Sentiment is

stationary, most of the cryptocurrency prices exhibit non-

stationary behavior, requiring second-order difference for

The outcome of the DCC-GARCH model is presented in

Table 7, illustrating the dynamic adjustments in conditional

correlation within a multivariate DCC model’s framework,

complemented by the GARCH model’s volatility insights.

These statistical results reveal that the price of Bitcoin and

news sentiment generally exhibit a similar directional

movement; however, this relationship is notably weak.

The

ρ

(Rho) value of 0.1145 indicates a low long-term

correlation between Bitcoin prices and news sentiment,

suggesting that, on average, they do not move together

closely. The

α

(Alpha) value is 0.00107, which is very small,

implying that recent news events exert minimal in

the immediate volatility of Bitcoin’s price. This means that

uence on

newinformationorshocksfromnewshaveanegligibleshort-

term impact on Bitcoin’s volatility.

Conversely, the

β

(Beta) value is 0.9874, which is quite

high, indicating that past volatility trends have a substantial

and enduring impact on the volatility of Bitcoin’s price. This

high Beta value suggests that historical news patterns are a

cant factor in the longer-term volatility of Bitcoin.

Transfer entropy values are calculated and detailed in

Table 8 and Table 9. It’s important to clarify that these

values should not be confused with directional or signal

relationships typical of correlations or coef

they should be understood as transfer entropy measures

cients. Instead,

owing from the ‘Sender’ to the ‘Receiver’, indicative of

the information transfer between the two entities. Our

ndings highlight signi

cant spillover effects within the

cryptocurrency markets, as gauged by the Transfer entropy

method.

Notably, cryptocurrencies with smaller market capital-

ization tend to react more sensitively compared to their

larger counterparts. For instance, XRP (ranked 7th) and

ADA (ranked 10th) emerge as the most notable recipients.

As shown in Table 9, they are the most sensitive to changes,

receiving signals from 7 other sources, indicating their high

reactivity to market information including news sentiment.

BNB, on the other hand, sends signals to 7 other cryptocur-

rencies, showcasing its role as a signi

Conversely, cryptocurrencies with the largest market

capitalization exhibit lower levels of information exchange.

For example, BTC sends information to 7 other cryptocur-

rencies, illustrating its central role in the market. Despite its

uence, BTC receives only 4 signals, re

its relative stability and lower sensitivity to external shocks

ecting

compared to smaller cryptocurrencies.

News events that substantially affect Bitcoin’s valuation

often initiate a domino effect, impacting the valuations of

other cryptocurrencies. For instance, it has been observed

that news impacts the prices of BNB, ETH, and XRP. As

shown in Table 9, News Sentiment sends 4 shocks to

othercryptocurrenciesbutreceivesonly1shockfromanother

cryptocurrency, highlighting its role as a signi

Nonetheless, Bitcoin exerts a more profound in

these cryptocurrencies. Given that Bitcoin accounts for

uence on

approximately 50% of the total market capitalization of all

cryptocurrencies, our observations align with those reported

inthestudybyZhangetal.[83].Thissigni

underscores Bitcoin’s extensive connectivity and in

cantmarketshare

over other cryptocurrencies, including BNB, ETH, DOGE,

uence

TRON, XRP, SOL and ADA.

ndings of this case study reveal a discernible but

modest correlation between news sentiment and Bitcoin

price

uctuations. Utilizing the robust FinBERT [12] model

for sentiment analysis and the DCC-GARCH technique for

nancial analysis, we gleaned signi

dynamic interplay between public sentiment, as re

cant insights into the

media, and Bitcoin’s price volatility. Speci

ected in

tical results from the DCC-GARCH model suggested that

cally, the statis-

historical news patterns wield a more substantial impact on

Bitcoin’s longer-term volatility than immediate news events.

These

ndings provide insights into the interrelationships

betweennewsandBitcoinprice,underscoringtheimportance

of monitoring news for cryptocurrency.

Thisinvestigationissubjecttocertainconstraints.Notably,

the scope of the data utilized could be more extensive

regarding the timeframe and the range of currencies

examined, which may in

uence the perceived relationship

between the variables. Including more data and additional

cryptocurrencies in future analyses could alter the outcomes of

this study. Furthermore, the diversity of methodologies

comparing their results. Exploring factors in

tocurrency development is an evolving area of academic

uencing cryp-

interest that warrants further exploration. Future studies aim to

overcome these limitations by incorporating broader and

more varied datasets and adopting more uniform research

methods, contributing to a more cohesive understanding

among scholars in the

This section delves into the multifaceted challenges and

limitations of using LLMs in sentiment analysis. The techni-

cal dif

culties are paramount, highlighted by the signi

computational and storage demands of evolving models like

cant

GPT-1 [46] to those with trillions of parameters, raising

concerns about accessibility in resource-limited contexts.

LLMsalsoneedhelpwithgeneralizability,oftenneedinghelp

maintaining consistent performance across diverse domains

and tasks. This points to a need for models that are more

adaptable and versatile. Additionally, the interpretability and

ethical usage of LLMs are crucial, especially in critical

sectors like

nance, where the opaque nature of these models

can hinder trust and reliability.

The evolution of LLMs has been characterized by a

substantial increase in their size, with a progression from

GPT-1’s 117 million parameters [46] to GPT-2’s 1.5 billion

[45] and a dramatic leap to GPT-3’s 175 billion parame-ters

[47]. More recent models have continued this trend,

reaching into the trillions of parameters [84]. Such vast

sizes present formidable challenges regarding storage, mem-

ory, and computational requirements. These challenges are

particularly acute in scenarios with limited resources or real-

time demands, especially when developers cannot access

high-powered GPUs or TPUs. For instance, FinBERT is a

pre-trained model with 110 million parameters, resulting in

a considerable size of 438 MB [12]. The Hugging Face

team [85] notes that training a 176 billion parameter model

like BLOOM [86] on a 1.5 TB dataset consumes 1,082,880

[87] on the Pile dataset [88], which includes over 825 GiB

of raw text data, requires eight NVIDIA A100-SXM4-40GB

GPUs. This extensive training can last up to 1,830 hours or

approximately 76 days.

Beyond the monetary costs, these models also incur

cant energy expenses. Predictions indicate a massive

increase in energy usage by platforms employing LLMs [89],

raising environmental concerns. However, a growing body of

research is aimed at mitigating these challenges. For

example, Wang et al. [90] have demonstrated a distillation

method that successfully compresses the MiniLM model to a

mere 66 million parameters, signi

cantly reducing its size

while maintaining ef

ciency. Increasing LLM sizes poses a

complex challenge, necessitating ongoing efforts for more

ef

cient deployment strategies.

Generalizability in LLMs pertains to their capability to

perform tasks accurately and consistently across various

domains, datasets, or functions that differ from their initial

training environment. Although LLMs are often trained on

extensive datasets, encompassing a broad range of

knowledge, their ef

cacy can be less reliable when applied to

unique or niche tasks outside their primary training scope. This

limitation becomes evident in diverse applications, from

coding projects to document analysis, where the context and

semantics can vary signi

cantly across different projects,

languages, or domains.

To enhance the generalizability of LLMs, it is crucial to

engage in meticulous

ne-tuning, apply rigorous validation

across diverse datasets, and establish continuous feedback

mechanisms. These steps are vital to prevent models from

becoming overly specialized in their training data, which

can severely restrict their applicability in various real-world

scenarios. However, despite these precautions, recent studies

indicate that LLMs often need help to extend their high-

performance levels to inputs markedly different from their

training data [91]. This limitation highlights a signi

in the current capabilities of LLMs. The challenge, therefore,

cant gap

lies in developing LLMs that possess extensive knowledge

and understanding gleaned from large datasets and exhibit

the

exibilityandadaptabilityrequiredtofunctioneffectively

across a wide range of contexts. Addressing this challenge

involvesre

ningthetrainingprocessandinnovatinginmodel

architecture and learning algorithms.

Interpretability and trustworthiness are pivotal in integrating

LLMs for sentiment analysis tasks. The primary challenge

lies in demystifying the decision-making processes of these

models. Due to their ‘black-box’ nature, elucidating the

mechanisms through which they discern sentiment from text is

often challenging. Recent studies [92] have underscored this

issue, revealing that while LLMs are pro

sentiment analysis, their opaque internal workings remain a

cient in

cant barrier. This obscurity in understanding how these

models arrive at their conclusions can generate apprehension

and reluctance among users, particularly investors who rely

on clear and logical reasoning for decision-making [93].

Investors may only trust the outputs of LLMs with a

transparent understanding of the underlying processes.

To foster trust in LLMs, it is essential to develop and

implement techniques and tools that shed light on the

internal mechanics of these models. Such efforts would

enable developers and users to trace and understand the

rationale behind the outputs generated by LLMs. Improving

interpretability and trustworthiness is a technical necessity

and a step towards broader acceptance and use of LLMs in

sentiment analysis, leading to more ef

Another aspect contributing to the challenge is the

closed nature of many LLMs. Often, it needs to be more

transparent about what data these models have been trained

on, raising questions about the source training data’s quality,

representativeness, and ownership. This lack of transparency

extends to concerns over the ownership of derivative data

produced by the models [95]. Furthermore, the potential

vulnerability of LLMs to various adversarial attacks, where

inputs are maliciously designed to manipulate or confuse

the models, adds another layer of complexity. These risks

emphasize the need for robust security measures and ethical

considerations in developing and deploying LLMs.

This section highlights the future opportunities for LLMs in

sentiment analysis. As these models evolve and gain

prominence in academic research, we explore the emerging

trends and potential advancements that could shape their role in

sentiment analysis. This section re

ciency and effectiveness, expands their

naturallanguageprocessingcapabilitiestoencompassamore

comprehensive array of input forms, and discusses enhancing

their performance in existing sentiment analysis tasks.

ascent of ChatGPT in academic research highlights its

growing prominence and acceptance in scholarly circles.

Researchers have increasingly favored ChatGPT over other

LLMs and their applications since its release, primarily due

to its computational ef

ciency, versatility in handling diverse

tasks, and potential for cost-effectiveness [96]. Beyond its

application in sentiment analysis, ChatGPT has spearheaded

an era of enhanced collaboration in the

nancial sector.

This trend marks a signi

cant shift towards incorporating

sophisticated natural language understanding into sentiment

analysis[97].Byexaminingtheseevolvingdynamics,wecan

anticipate the future trajectory of LLMs like ChatGPT in

re

ning and revolutionizing sentiment analysis processes.

These developments indicate the transformative potential of

LLMs in sentiment analysis.

Regarding the utilization of LLMs, the choice between

using commercially available pre-trained models like GPT-4

and opting for open-source alternatives such as LLaMA [51],

LlaMA 2 [52], and Alpaca

7

presents distinct avenues for

7

https://github.com/tatsu-lab/stanford_alpaca

customization in specialized tasks. The critical difference

between these approaches lies in their level of control and

personalization. Despite their proprietary nature, pre-trained

models like GPT-4 enable quick, task-speci

with minimal data requirements. This approach reduces

c adaptations

computational demands and expedites deployment. In

contrast, open-source frameworks like LLaMA provide a

foundation for extensive tailoring. While these models arrive

pre-trained, they can be further adapted, with organizations

often modifying and retraining them on large-scale datasets

speci

c to their needs [98]. Although this process demands

substantial computational resources and investment, it allows

for creating models intricately tailored to speci

Throughout our analysis, it became apparent that most data

inputs for LLMs in sentiment analysis were text-based. This

nding aligns with traditional NLP approaches, yet there

needs to be a noticeable gap in utilizing more diverse and

complex datasets, particularly graph-based ones. Embracing a

more comprehensive array of natural language inputs, such as

spoken language, diagrams, and multimodal data, could

cantly expand the capabilities of LLMs in capturing

and interpreting varied forms of user sentiment [99].

IntegratingspokenlanguageintoLLMscouldenhanceuser

interactions, enabling the models to process more natural and

contextually rich conversations. This addition would allow

LLMs to understand better nuances in tone, intonation, and

colloquial expressions, which often need to be improved in

text-based communication. Similarly, including diagrams

could provide valuable visual representations of complex

ideas or emotions, offering a unique dimension to sentiment

analysis [100]. Diagrams can be a powerful tool to convey

information that may be dif

cult to express through words

alone.

Moreover, multimodal inputs that amalgamate text, audio,

and visual elements could lead to a more holistic under-

standing of context. Such a comprehensive approach would

likely result in more accurate and context-sensitive senti-

ment analysis outcomes. For instance, combining textual

data with vocal intonations and facial expressions could

better understand the user’s emotional state and intentions

In academic research, establishing a universal and adaptable

evaluation framework for LLMs in sentiment analysis is

becoming increasingly imperative. Such a framework is

essential for conducting systematic and consistent assess-

ments of LLMs, focusing on their performance, ef

and potential limitations. This standardization would serve as

cacy,

a critical benchmark, enabling researchers to verify the

practical readiness of these models for various applications.

A standardized evaluation framework would offer a compre-

hensive set of criteria and metrics against which to measure

LLMs, ensuring that their capabilities are accurately and

objectively assessed [103].

In academia, where rigorous analysis and validation are

paramount, the absence of such a framework can lead to

fragmentedandinconsistentevaluationsofLLMs,potentially

impeding their development and adoption. By establishing a

universally accepted framework, researchers can compare

different LLMs on a level playing

understanding of each model’s strengths and areas for

eld, fostering a clearer

improvement. This framework should ideally encompass a

range of considerations, including accuracy in sentiment

detection, adaptability to different linguistic contexts, com-

putational ef

ciency, and ethical concerns such as bias and

fairness [104], [105].

Furthermore, a universal evaluation framework would

facilitate responsible LLM adoption in academic research

[106]. It would provide scholars with the tools to decide

which models best suit their research needs and objectives.

In this comprehensive literature review, we adeptly examine

the intersection of LLMs and sentiment analysis within

nancial markets, providing a detailed exploration of

LLMs’ evolution, application, and future opportunities in

this domain. The review navigates through the intrica-cies

of sentiment analysis, underlining its signi

understanding market dynamics and investor behavior. Our

cance in

meticulousanalysisofLLMs,mainlytheirdevelopmentfrom

BERT [14] to more sophisticated models like FinBERT [12]

and ChatGPT, reveals these models’ substantial impact on

nancial sentiment analysis.

The review methodically dissects the role of LLMs in

various

nancial contexts, from cryptocurrency market pre-

diction to stock price forecasting, showcasing their capability

to extract and interpret complex economic sentiments. The

case study on Bitcoin price and news sentiment further

exempli

es the practical application of LLMs, reinforcing

that sentiment analysis, powered by advanced language

models, is pivotal in deciphering market trends.

However, the review is open to addressing the challenges

and limitations inherent in the current state of LLMs. Issues

such as the immense computational requirements, dif

in generalizability and interpretability, and ethical concerns

culties

are thoughtfully discussed, providing a balanced perspective.

We call for more ef

cient deployment strategies, improved

generalizability, and enhanced interpretability is particularly

compelling, indicating the need for continued innovation in

Looking to the future, integrating more diverse data types

and establishing a universal evaluation framework are essen-

tialstepstowardenhancingtheef

analysis. The potential expansion of LLM capabilities to

cacyofLLMsinsentiment

include multimodal data inputs and the implementation of a

standard evaluation framework are highlighted as promising

avenues for research and development.

[1] M. Baker and J. Wurgler, ‘‘Investor sentiment in the stock market,’’

J. Econ. Perspect., vol. 21, no. 2, pp. 129–152, 2007.

[2] P. C. Tetlock, ‘‘Giving content to investor sentiment: The role of

media in the stock market,’’ J. Finance, vol. 62, pp. 1139–1168, Jun.

2007. [Online]. Available: https://onlinelibrary.wiley.com/doi/full/

10.1111/j.1540-6261.2007.01232.x

[3] L. A. Smales, ‘‘The importance of fear: Investor sentiment and

stock market returns,’’ Appl. Econ., vol. 49, no. 34, pp. 3395–3421,

Jul. 2017. [Online]. Available: https://www.tandfonline.com/doi/abs/

10.1080/00036846.2016.1259754

[4] T. Rao and S. Srivastava. (2012). Analyzing Stock Market Movements

Using Twitter Sentiment Analysis. [Online]. Available: http://dx.doi.org/

10.1109/ASONAM.2012.30 and https://repository.lincoln.ac.uk/articles/

conference_contribution/Analyzing_stock_market_movements_using_T

witter_sentiment_analysis/25165223/2?

[5] E. Cambria and B. White, ‘‘Jumping NLP curves: A review of natural

le=44450105

language processing research,’’ IEEE Comput. Intell. Mag., vol. 9, no. 2,

pp. 48–57, May 2014.

[6] V. Ramiah, X. Xu, and I. A. Moosa, ‘‘Neoclassical

nance and noise traders: A review and assessment of the literature,’’ Int.

Rev. Financial Anal., vol. 41, pp. 89–100, Oct. 2015.

[7] F. Wu, Y. Huang, and Y. Song, ‘‘Structured microblog sentiment clas-

si

cation via social context regularization,’’ Neurocomputing, vol. 175,

pp. 599–609, Jan. 2016.

[8] T. Al-Moslmi, S. Gaber, M. Albared, and N. Omar. (2016). Feature

Selection Methods Effects on Machine Learning Approaches in Malay

Sentiment Analysis. [Online]. Available: https://www.researchgate.

net/publication/308968243

[9] R. C. Moore and W. Lewis, ‘‘Intelligent selection of language model

training data,’’ in Proc. ACL Conf. Short Papers, 2010, pp. 220–224.

[10] S. Wu, O. Irsoy, S. Lu, V. Dabravolski, M. Dredze, S. Gehrmann, P.

Kambadur, D. Rosenberg, and G. Mann, ‘‘BloombergGPT: A large

language model for

nance,’’ 2023, arXiv:2303.17564.

[11] K. Mishev, A. Gjorgjevikj, I. Vodenska, L. T. Chitkushev, and D.

Trajanov, ‘‘Evaluation of sentiment analysis in

transformers,’’ IEEE Access, vol. 8, pp. 131662–131682, 2020.

nance: From lexicons to

[12] D. Araci, ‘‘FinBERT: Financial sentiment analysis with pre-trained

language models,’’ 2019, arXiv:1908.10063.

[13] P. Seroyizhko, Z. Zhexenova, M. Z. Sha

q, F. Merizzi, A. Galassi, and F.

Ruggeri, ‘‘A sentiment and emotion annotated dataset for Bitcoin price

forecasting based on Reddit posts,’’ in Proc. 4th Workshop Financial

Technol. Natural Lang. Process. (FinNLP), 2022, pp. 203–210. [Online].

Available: https://aclanthology.org/2022.

[14] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, ‘‘BERT: Pre-training of

nnlp-1.27

deep bidirectional transformers for language understanding,’’ in Proc.

Conf. North Amer. Chapter Assoc. Comput. Linguistics, Hum. Lang.

Technol., vol. 1, Oct. 2018, pp. 4171–4186.

[15] J. Kocon, I. Cichecki, O. Kaszyca, M. Kochanek, D. Szydlo, J. Baran,

J. Bielaniewicz, M. Gruza, A. Janz, K. Kanclerz, A. Kocon, B. Koptyra,

W. Mieleszczenko-Kowszewicz, P. Milkowski, M. Oleksy, M. Piasecki,

L. Radlinski, K. Wojtasik, S. Wozniak, and P. Kazienko, ‘‘ChatGPT:

Jack of all trades, master of none,’’ Inf. Fusion, vol. 99, Nov. 2023,

Art. no. 101861.

[16] M. Chakraborty and S. Subramaniam, ‘‘Does sentiment impact

cryptocurrency?’’ J. Behav. Finance, vol. 24, no. 2, pp. 202–218,

Apr. 2023. [Online]. Available: https://www.tandfonline.com/doi/abs/

10.1080/15427560.2021.1950723

[17] A. H. Huang, H. Wang, and Y. Yang, ‘‘FinBERT: A large language model

for extracting information from

nancial text,’’ Contemp. Accounting

Res., vol. 40, no. 2, pp. 806–841, May 2023. [Online]. Available:

https://onlinelibrary.wiley.com/doi/full/10.1111/1911-3846.12832

[18] H. Tong, J. Li, N. Wu, M. Gong, D. Zhang, and Q. Zhang, ‘‘Ploutos:

Towards interpretable stock movement prediction with

language model,’’ 2024, arXiv:2403.00782.

nancial large

[19] A. S. George and A. H. George, ‘‘A review of ChatGPT AIs impact on

several business sectors,’’ Partners Universal Int. Innov. J., vol. 1, no. 1,

[20] N. A. Sharma, A. B. M. S. Ali, and M. A. Kabir, ‘‘A review of sentiment

analysis: Tasks, applications, and deep learning techniques,’’ Int. J. Data

Sci. Anal., pp. 1–38, Jul. 2024, doi:

10.1007/s41060-024-00594-x.

[21] M. A. K. Raiaan, M. S. H. Mukta, K. Fatema, N. M. Fahad, S. Sakib, M.

M. J. Mim, J. Ahmad, M. E. Ali, and S. Azam, ‘‘A review on large

language models: Architectures, applications, taxonomies, open issues

and challenges,’’ IEEE Access, vol. 12, pp. 26839–26874, 2024.

[22] B. Chen, Z. Wu, and R. Zhao, ‘‘From

f

i

ction to fact: The growing role of

generative AI in business and

f

i

nance,’’ J. Chin. Econ. Bus. Stud., vol. 21,

no. 4, pp. 471–496, Oct. 2023.

[23] M. M. Dong, T. C. Stratopoulos, and V. X. Wang, A Scoping Review of

ChatGPTResearchinAccountingandFinance,T.C.WangandV.Xiaoqi,

Eds., Dec. 2023. [Online]. Available: https://ssrn.com/abstract=4680203

and http://dx.doi.org/10.2139/ssrn.4680203

[24] S. A. Farimani, M. V. Jahan, and A. M. Fard, ‘‘From text representation to

f

i

nancial market prediction: A literature review,’’ Information, vol. 13, no.

10, p. 466, Sep. 2022.

[25] A. Koshiyama, N. Firoozye, and P. Treleaven, ‘‘Algorithms in future

capital markets: A survey on AI, ML and associated algorithms in capital

markets,’’ in Proc. 1st ACM Int. Conf. AI Finance, 2020, pp. 1–8.

[26] O. Bashchenko, ‘‘Bitcoin price factors: Natural language processing

approach,’’ SSRN Electron. J., vol. 13, pp. 22–48, Mar. 2022. [Online].

Available: https://papers.ssrn.com/abstract=4079091

[27] B. N. Thanh, A. T. Nguyen, T. T. Chu, and S. Ha. (2023). ChatGPT,

Twitter Sentiment and Bitcoin Return. [Online]. Available: https://

papers.ssrn.com/abstract=4628097

[28] B. Kitchenham. (2007). Guidelines for Performing Systematic

Literature Reviews in Software Engineering. [Online]. Available:

https://www.researchgate.net/publication/302924724

[29] B. Kitchenham, L. Madeyski, and D. Budgen, ‘‘SEGRESS: Software

engineering guidelines for REporting secondary studies,’’ IEEE Trans.

Softw. Eng., vol. 49, no. 3, pp. 1273–1298, Mar. 2023.

[30] H. Zhao, Z. Liu, Z. Wu, Y. Li, T. Yang, P. Shu, S. Xu, H. Dai, L. Zhao, G.

Mai, N. Liu, and T. Liu, ‘‘Revolutionizing

f

i

nance with LLMs: An

overview of applications and insights,’’ 2024, arXiv:2401.11641.

[31] K. Du, F. Xing, R. Mao, and E. Cambria, ‘‘Financial sentiment analysis:

Techniques and applications,’’ ACM Comput. Surv., vol. 56, no. 9, pp.

1–42, Oct. 2024.

[32] M. N. Ashtiani and B. Raahemi, ‘‘News-based intelligent prediction of

f

i

nancial markets using text mining and machine learning: A

systematic literature review,’’ Expert Syst. Appl., vol. 217, May 2023,

Art. no. 119509.

[33] M. Shanahan, ‘‘Talking about large language models,’’ 2022,

arXiv:2212.03551.

[34] J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi,

Q. V. Le, and D. Zhou, ‘‘Chain-of-thought prompting elicits reasoning in

large language models,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 35,

2022, pp. 24824–24837.

[35] R. Taylor, M. Kardas, G. Cucurull, T. Scialom, A. Hartshorn, E. Saravia, A.

Poulton, V. Kerkez, and R. Stojnic, ‘‘Galactica: A large language model

for science,’’ 2022, arXiv:2211.09085.

[36] J. Hoffmann et al., ‘‘Training compute-optimal large language models,’’

2022, arXiv:2203.15556.

[37] J. Xu Zhao, Y. Xie, K. Kawaguchi, J. He, and M. Q. Xie, ‘‘Automatic

model selection with large language models for reasoning,’’ 2023,

arXiv:2305.14333.

[38] S. Pan, L. Luo, Y. Wang, C. Chen, J. Wang, and X. Wu, ‘‘Unifying

large language models and knowledge graphs: A roadmap,’’ 2023,

arXiv:2306.08302.

[39] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut,

‘‘ALBERT: A lite BERT for self-supervised learning of language

representations,’’ in Proc. 8th Int. Conf. Learn. Represent. (ICLR), 2020,

pp. 1–17.

[40] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L.

Zettlemoyer, and V. Stoyanov, ‘‘RoBERTa: A robustly optimized

BERT pretraining approach,’’ 2019, arXiv:1907.11692.

[41] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L.

Kaiser, and I. Polosukhin, ‘‘Attention is all you need,’’ in Proc. Adv.

Neural Inf. Process. Syst., vol. 30, 2017, pp. 1–11.

[42] M. Kulakowski and F. Frasincar, ‘‘Sentiment classi

f

i

cation of

cryptocurrency-related social media posts,’’ IEEE Intell. Syst., vol. 38,

no. 4, pp. 5–9, Jul. 2023.

[43] Y. Yu, H. Li, Z. Chen, Y. Jiang, Y. Li, D. Zhang, R. Liu, J. W. Suchow, and K.

Khashanah, ‘‘FinMem: A performance-enhanced LLM trading agent with

layered memory and character design,’’ 2023, arXiv:2311.13743.

[44] Y. Li, Y. Yu, H. Li, Z. Chen, and K. Khashanah, ‘‘TradingGPT: Multi-

agent system with layered memory and distinct characters for enhanced

f

i

nancial trading performance,’’ 2023, arXiv:2309.03736.

[45] A.Radford,J.Wu,R.Child,D.Luan,D.Amodei,andI.Sutskever,‘‘Lan-

guage models are unsupervised multitask learners,’’ OpenAI Blog, vol. 1,

no. 8, p. 9, 2019. [Online]. Available: https://github.com/codelucas/

newspaper

[46] A. Radford, K. Narasimhan, T. Salimans, and I. Sutskever, Improving

Language Understanding by Generative Pre-training. Accessed: Feb. 3,

2024. [Online]. Available: https://gluebenchmark.com/leaderboard

[47] T. B. Brown et al., ‘‘Language models are few-shot learners,’’ in

Proc. NIPS, 2020, pp. 1877–1901. [Online]. Available: https://

commoncrawl.org/the-data/

[48] J. Achiam et al., ‘‘GPT-4 technical report,’’ 2023, arXiv:2303.08774.

[49] G. Fatouros, J. Soldatos, K. Kouroumali, G. Makridis, and D. Kyriazis,

‘‘Transforming sentiment analysis in the

f

i

nancial domain with

ChatGPT,’’ Mach. Learn. Appl., vol. 14, Dec. 2023, Art. no. 100508.

[50] B.Zhang,H.Yang,andX.-Y.Liu,‘‘Instruct-FinGPT:Financialsentiment

analysisbyinstructiontuningofgeneral-purposelargelanguagemodels,’’

2023, arXiv:2306.12659.

[51] H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux,

T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez,

A. Joulin, E. Grave, and G. Lample, ‘‘LLaMA: Open and ef

f

i

cient

foundation language models,’’ 2023, arXiv:2302.13971.

[52] H. Touvron et al., ‘‘Llama 2: Open foundation and

f

i

ne-tuned chat

models,’’ 2023, arXiv:2307.09288.

[53] Q. Xie, W. Han, X. Zhang, Y. Lai, M. Peng, A. Lopez-Lira, and J. Huang,

‘‘PIXIU: A large language model, instruction data and evaluation

benchmark for

f

i

nance,’’ 2023, arXiv:2306.05443.

[54] B. Peng, E. Chersoni, Y.-Y. Hsu, L. Qiu, and C.-R. Huang, ‘‘Supervised

cross-momentum contrast: Aligning representations with prototypical

examples to enhance

f

i

nancial sentiment analysis,’’ Knowl.-Based Syst.,

vol. 295, Jul. 2024, Art. no. 111683.

[55] C. He, C. Li, T. Han, and L. Shen, ‘‘Assessing and enhancing LLMs: A

physics and history dataset and one-more-check pipeline method,’’ in

Proc. Int. Conf. Neural Inf. Process., 2024, pp. 504–517.

[56] Z.Chen,W.Chen,C.Smiley,S.Shah,I.Borova,D.Langdon,R. Moussa,

M. Beane, T.-H. Huang, B. Routledge, and W. Y. Wang, ‘‘FinQA:

A dataset of numerical reasoning over

f

i

nancial data,’’ in Proc. Conf.

Empirical Methods Natural Lang. Process., 2021, pp. 3697–3711.

[57] Z. Liu, D. Huang, K. Huang, Z. Li, and J. Zhao, ‘‘FinBERT: A pre-trained

f

i

nancial language representation model for

f

i

nancial text mining,’’ in

Proc. 29th Int. Joint Conf. Artif. Intell., Jul. 2020, pp. 4513–4519.

[Online]. Available: http://commoncrawl.org/

[58] P. Malo, A. Sinha, P. Korhonen, J. Wallenius, and P. Takala, ‘‘Good debt or

bad debt: Detecting semantic orientations in economic texts,’’ J. Assoc. Inf.

Sci. Technol., vol. 65, no. 4, pp. 782–796, Apr. 2014.

[59] K. Cortis, A. Freitas, T. Daudert, M. Huerlimann, M. Zarrouk, S.

Handschuh, and B. Davis, ‘‘SemEval-2017 task 5: Fine-grained

sentiment analysis on

f

i

nancial microblogs and news,’’ in Proc. 11th Int.

Workshop Semantic Eval. (SemEval), 2017, pp. 519–535.

[60] G. Kim, M. Kim, B. Kim, and H. Lim, ‘‘CBITS: Crypto BERT incorpo-

rated trading system,’’ IEEE Access, vol. 11, pp. 6912–6921, 2023.

[61] Y. Zou and D. Herremans, ‘‘PreBit—A multimodal model with Twitter

FinBERT embeddings for extreme price movement prediction of

Bitcoin,’’ Expert Syst. Appl., vol. 233, Dec. 2023, Art. no. 120838.

[62] A. Raheman, A. Kolonin, I. Fridkins, I. Ansari, and M. Vishwas, ‘‘Social

media sentiment analysis for cryptocurrency market prediction,’’ 2022,

arXiv:2204.10185.

[63] M. Ortu, N. Uras, C. Conversano, S. Bartolucci, and G. Destefanis,

‘‘On technical trading and social media indicators for cryptocurrency

price classi

f

i

cation through deep learning,’’ Expert Syst. Appl., vol. 198,

Jul. 2022, Art. no. 116804.

[64] B. Fazlija and P. Harder, ‘‘Using

f

i

nancial news sentiment for stock price

direction prediction,’’ Mathematics, vol. 10, no. 13, p. 2156, Jun. 2022.

[Online]. Available: https://www.mdpi.com/2227-7390/10/13/2156/htm

[65] U. Gupta, ‘‘GPT-InvestAR: Enhancing stock investment strategies

through annual report analysis with large language models,’’ 2023,

arXiv:2309.03079.

V

OLU

M

E

12

,

2024

134059

[66] X. Deng, V. Bashlovkina, F. Han, S. Baumgartner, and M. Bendersky,

‘‘What do LLMs know about

nancial markets? A case study on Reddit

market sentiment analysis,’’ 2022, arXiv:2212.11311.

[67] H. Q. Abonizio, E. C. Paraiso, and S. Barbon, ‘‘Toward text data

augmentation for sentiment analysis,’’ IEEE Trans. Artif. Intell., vol. 3,

no. 5, pp. 657–668, Oct. 2022.

[68] D. Ider and S. Lessmann, ‘‘Forecasting cryptocurrency returns from sen-

timent signals: An analysis of BERT classi

ers and weak supervision,’’

2022, arXiv:2204.05781.

[69] J. de Curtò, I. de Zarzà, G. Roig, J. C. Cano, P. Manzoni, and C. T.

Calafate, ‘‘LLM-informed multi-armed bandit strategies for non-

stationaryenvironments,’’Electronics,vol.12,no.13,p. 2814,Jun.2023.

[Online]. Available: https://www.mdpi.com/2079-9292/12/13/2814/htm

[70] M. Fernandes, S. Khanna, L. Monteiro, A. Thomas, and G. Tripathi,

‘‘Bitcoin price prediction,’’ in Proc. Int. Conf. Adv. Comput., Commun.,

Control (ICAC3), Dec. 2021, pp. 1–4.

[71] P. Malo, A. Sinha, P. Takala, P. Korhonen, and J. Wallenius, ‘‘Good debt or

bad debt: Detecting semantic orientations in economic texts,’’ 2013,

arXiv:1307.5336.

[72] S. A. Farimani, M. V. Jahan, A. M. Fard, and S. R. K. Tabbakh,

‘‘Investigating the informativeness of technical indicators and news

sentiment in

nancial market price prediction,’’ Knowl.-Based Syst.,

vol. 247, Jul. 2022, Art. no. 108742.

[73] T. C. Chiang, B. N. Jeon, and H. Li, ‘‘Dynamic correlation analysis of

nancial contagion: Evidence from Asian markets,’’ J. Int. Money

Finance, vol. 26, no. 7, pp. 1206–1228, Nov. 2007.

[74] K. J. Forbes and R. Rigobon, ‘‘No contagion, only interdependence:

Measuring stock market comovements,’’ J. Finance, vol. 57,

no. 5, pp. 2223–2261, Oct. 2002. [Online]. Available: https://

onlinelibrary.wiley.com/doi/full/10.1111/0022-1082.00494

[75] L. Barnett, A. B. Barrett, and A. K. Seth, ‘‘Granger causality and transfer

entropy are equivalent for Gaussian variables,’’ Phys. Rev. Lett., vol. 103,

no. 23, Dec. 2009, Art. no. 238701.

[76] C. E. Shannon, ‘‘A mathematical theory of communication,’’ Bell Syst.

Tech. J., vol. 27, no. 3, pp. 379–423, Jul. 1948.

[77] S. Kullback and R. A. Leibler, ‘‘On information and suf

Math. Statist., vol. 22, no. 1, pp. 79–86, 1951.

ciency,’’ Ann.

[78] T. Schreiber, ‘‘Measuring information transfer,’’ Phys. Rev. Lett., vol. 85,

no. 2, pp. 461–464, Jul. 2000.

[79] T. Dimp

and F. J. Peter, ‘‘Using transfer entropy to measure information

nancial markets,’’ Stud. Nonlinear Dyn. Econometrics,

vol. 17, no. 1, pp. 85–102, 2013.

[80] S. Bekiros, D. K. Nguyen, L. S. Junior, and G. S. Uddin, ‘‘Information

diffusion, cluster formation and entropy-based network dynamics in

equity and commodity markets,’’ Eur. J. Oper. Res., vol. 256, no. 3, pp.

945–961, Feb. 2017.

[81] T. A. Brown, Confirmatory Factor Analysis for Applied Research. NY,

USA: Guilford publications, 2015.

[82] P. Wollstadt, M. Martínez-Zarzuela, R. Vicente, F. J. Díaz-Pernas, and M.

Wibral, ‘‘Ef

cient transfer entropy analysis of non-stationary neural time

series,’’ PLoS ONE, vol. 9, no. 7, Jul. 2014, Art. no. e102833.

[83] H. Zhang, H. Hong, Y. Guo, and C. Yang, ‘‘Information spillover effects

frommediacoveragetothecrudeoil,gold,andBitcoinmarketsduringthe

COVID-19 pandemic: Evidence from the time and frequency domains,’’

Int. Rev. Econ. Finance, vol. 78, pp. 267–285, Mar. 2022.

[84] S.Moss,‘‘Googlebrainunveilstrillion-parameterAIlanguagemodel,the

largest yet,’’ Tech. Rep., 2021.

[85] S. Bekman, ‘‘The technology behind Bloom training,’’ Tech. Rep., 2022.

[86] T. L. Scao et al., ‘‘BLOOM: A 176B-parameter open-access multilingual

language model,’’ 2022, arXiv:2211.05100.

[87] S. Black, S. Biderman, E. Hallahan, Q. Anthony, L. Gao, L. Golding,

H. He, C. Leahy, K. McDonell, J. Phang, M. Pieler, U. S. Prashanth,

S. Purohit, L. Reynolds, J. Tow, B. Wang, and S. Weinbach,

‘‘GPT-NeoX-20B: An open-source autoregressive language model,’’

2022, arXiv:2204.06745.

[88] L. Gao, S. Biderman, S. Black, L. Golding, T. Hoppe, C. Foster, J. Phang,

H. He, A. Thite, N. Nabeshima, S. Presser, and C. Leahy, ‘‘The Pile:

An 800 GB dataset of diverse text for language modeling,’’ 2021,

arXiv:2101.00027.

[89] M. C. Rillig, M. Ågerstrand, M. Bi, K. A. Gould, and U. Sauerland,

‘‘Risks and bene

ts of large language models for the environment,’’

Environ.Sci.Technol.,vol.57,no.9,pp. 3464–3466,Mar.2023.[Online].

Available: https://pubs.acs.org/doi/full/10.1021/acs.est.3c01106

[90] W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, and M. Zhou, ‘‘MiniLM:

Deep self-attention distillation for task-agnostic compression of pre-

trained transformers,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 2020,

2020, pp. 5776–5788.

[91] A. Albalak, A. Shrivastava, C. Sankar, A. Sagar, and M. Ross, ‘‘Data-

ef

ciency with a single GPU: An exploration of transfer methods for

small language models,’’ 2022, arXiv:2210.03871.

[92] X. Deng, V. Bashlovkina, F. Han, S. Baumgartner, and M. Bendersky,

‘‘What do LLMs Know about

nancial markets? A case study on Reddit

marketsentimentanalysis,’’inProc.ACMWebConf.,2022,pp. 107–110.

[Online]. Available: https://dl.acm.org/doi/10.1145/3543873.3587324

[93] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao,

‘‘ReAct: Synergizing reasoning and acting in language models,’’ 2022,

arXiv:2210.03629.

[94] X. Li, H. Xiong, X. Li, X. Wu, X. Zhang, J. Liu, J. Bian, and

D. Dou, ‘‘Interpretable deep learning: Interpretation, interpretability,

trustworthiness, and beyond,’’ Knowl. Inf. Syst., vol. 64, no. 12,

pp. 3197–3234, Dec. 2022. [Online]. Available: https://link.springer.

com/article/10.1007/s10115-022-01756-8

[95] S. Sinha, H. Chen, A. Sekhon, Y. Ji, and Y. Qi, ‘‘Perturbing inputs

for fragile interpretations in deep natural language pro-cessing,’’ in

Proc. 4th BlackboxNLP Workshop Analyzing Inter-preting Neural

Netw. NLP, 2021, pp. 420–434. [Online]. Available:

https://aclanthology.org/2021.blackboxnlp-1.33

[96] M. T. R. Laskar, M. S. Bari, M. Rahman, M. A. H. Bhuiyan, S. Joty,

and J. X. Huang, ‘‘A systematic study and comprehensive evaluation

of ChatGPT on benchmark datasets,’’ in Proc. Annu. Meeting Assoc.

Comput. Linguistics, 2023, pp. 431–469.

[97] M. U. Haque, I. Dharmadasa, Z. T. Sworna, R. N. Rajapakse, and H.

Ahmad, ‘‘‘I think this is the most disruptive technology’: Exploring

sentiments of ChatGPT early adopters using Twitter data,’’ 2022,

arXiv:2212.05856.

[98] I. Gur, O. Nachum, Y. Miao, M. Safdari, A. Huang, A. Chowdhery, S.

Narang, N. Fiedel, and A. Faust, ‘‘Understanding HTML with large

language models,’’ 2022, arXiv:2210.03945.

[99] X. Wang, J. He, Z. Jin, M. Yang, Y. Wang, and H. Qu, ‘‘M2Lens:

Visualizing and explaining multimodal models for sentiment analysis,’’

IEEE Trans. Vis. Comput. Graphics, vol. 28, no. 1, pp. 802–812, Jan.

2022.

[100] H. Song, J. Li, Z. Xia, Z. Yang, and X. Du, ‘‘Multimodal sentiment

analysis based on pre-LN transformer interaction,’’ in Proc. IEEE 6th

Inf. Technol. Mechatronics Eng. Conf. (ITOEC), vol. 6, Mar. 2022,

pp. 1609–1613.

[101] K. Dashtipour, M. Gogate, E. Cambria, and A. Hussain, ‘‘A novel

context-aware multimodal framework for Persian sentiment analysis,’’

Neurocomputing, vol. 457, pp. 377–388, Oct. 2021.

[102] U.Sehar,S.Kanwal,K.Dashtipur,U.Mir,U.Abbasi,andF.Khan,‘‘Urdu

sentiment analysis via multimodal data mining based on deep learning

algorithms,’’ IEEE Access, vol. 9, pp. 153072–153082, 2021.

[103] A. Oussous, F.-Z. Benjelloun, A. A. Lahcen, and S. Belfkih,

‘‘ASA: A framework for Arabic sentiment analysis,’’ J. Inf. Sci.,

vol. 46, no. 4, pp. 544–559, Aug. 2020. [Online]. Available:

https://journals.sagepub.com/doi/10.1177/0165551519849516

[104] Z. Ke, J. Sheng, Z. Li, W. Silamu, and Q. Guo, ‘‘Knowledge-guided

sentiment analysis via learning from natural language explanations,’’

IEEE Access, vol. 9, pp. 3570–3578, 2021.

[105] Q. Zhang, J. Zhou, Q. Chen, Q. Bai, J. Xiao, and L. He, ‘‘A

knowledge-enhanced adversarial model for cross-lingual structured

sentiment analysis,’’ in Proc. Int. Joint Conf. Neural Netw., Jul. 2022,

pp. 1–8.

[106] G. F. N. Mvondo, B. Niu, and S. Eivazinezhad, ‘‘Generative con-

versational AI and academic integrity: A mixed method investigation

to understand the ethical use of LLM chatbots in higher educa-

tion,’’ SSRN Electron. J., 2023. [Online]. Available: https://ssrn.com/

abstract=4548263 and http://dx.doi.org/10.2139/ssrn.4548263

software engineering from Jiangxi University of

Finance and Economics, in 2020. He is currently

pursuing the master’s degree with The Univer-

sity of Auckland. His research interests include

machine learning and large language model to

received the B.Tech. degree in information tech-

nology from Anna University, Chennai, India,

the M.Tech. degree in computer science and

engineering from Vellore Institute of Technology

(VIT), Vellore, India, and the Ph.D. degree from

the Faculty of Information and Communication

Engineering, Anna University, in 2023. He is an

Assistant Professor with the School of Computer

Science Engineering and Information Systems

(SCORE), VIT University. He is pro

cient with simulator tools MATLAB,

ns-3, Mininet, OpenNet VM, and P4 programming. He is exposed to open

source tools, such as OpenStack, Cloudify, OPNFV, and Cloud-Native

Computing Foundation (CNCF). His research interests include the cloud-

native deployment, SDN, NFV, 5G/6G networks, AI/ML based networking,

the Internet of Vehicles, and UAV communications.

(Member, IEEE) received the

M.Sc. degree in parallel and distributed com-

puting from Universiti Putra Malaysia, and the

Ph.D. degree in information technology from

the University of Tasmania, Australia. He is a

Senior Lecturer of information systems with

Queensland University of Technology (QUT).

He has authored more than 50 peer-reviewed

scienti

c research articles. His research interests

include distributed computing (fog/edge/cloud),

the Internet of Things (IoT), AI and ML, software-de